Skip to main content
NCBI home page
JAMA Network logoLink to JAMA Network
. 2026 Jan 28;9(1):e2555759. doi: 10.1001/jamanetworkopen.2025.55759

Deep-Learning Serial CT Prediction of Survival in Immunotherapy-Treated Non–Small Cell Lung Cancer

Chiharu Sako 1,✉, Brenda F Kurland 2, Taly G Schmidt 1, Jing He 3, Marius De Groot 4,5,6, Arpan A Patel 7, Dwight H Owen 8, Arya Amini 9, Brendan D Curti 10, Ronan J Kelly 11, Ray D Page 12, Aurelie Swalduz 13, Jean-Paul Beregi 14, Jan Chrusciel 15, Stephane Sanchez 15, Jakob Weiss 16, An Liu 9, Olivier Gevaert 17, Alma Hart 18, George R Simon 19, Ravi B Parikh 20, Jasper van der Aart 21, Petr Jordan 1
PMCID: PMC12853213  PMID: 41604152

Key Points

Question

Can a fully automated deep-learning radiomic biomarker based on serial computed tomography (CT) imaging improve prediction of overall survival in advanced non–small cell lung cancer (NSCLC) treated with immune checkpoint inhibitors?

Findings

In this prognostic study of 1830 patients with advanced NSCLC, serial CT response scores, using pretherapy and 12-week follow-up scans, had independent overall survival prognostic value in routine clinical practice and clinical trial multicenter validation datasets. Risk stratification compared favorably with conventional size-based tumor response assessments, including cohorts with stable disease per Response Evaluation Criteria in Solid Tumors.

Meaning

These findings suggest that serial CT response scores could improve clinical decision-making and enhance clinical trial designs for patients with NSCLC.

This prognostic study investigates whether an automated deep-learning imaging-basedbiomarker using pretherapy and follow-up computed tomography (CT) scans can improve prediction of overall survival in advanced non-small cell lung cancer.

Abstract

IMPORTANCE

Reliable early response biomarkers for overall survival (OS) are lacking for patients receiving immune checkpoint inhibitor (ICI) therapy for advanced non–small cell lung cancer (NSCLC). Existing imaging-based measures, such as Response Evaluation Criteria in Solid Tumors (RECIST) and tumor volume change (TVC), have limited predictive value for long-term outcomes, and advanced imaging-based biomarkers may enhance decision-making in clinical practice and clinical trials.

OBJECTIVE

To develop and validate a fully automated deep-learning imaging-based biomarker using pretherapy and 12-week follow-up computed tomography (CT) scans.

DESIGN, SETTING, AND PARTICIPANTS

This prognostic study used retrospectively collected HER data from routine clinical practice (RCP) and clinical trial data from 2013 to 2023. A model using serial CT scans, Serial CT response score (Serial CTRS) was developed using a RCP discovery dataset, validated with 10 US and European institution RCP test datasets, and independently validated on a multinational clinical phase 1 trial of dostarlimab (GARNET). Participants were adults with advanced NSCLC starting ICIs in the period from 2013 to 2021 (RCP discovery), from 2013 to 2022 (RCP test), or from 2017 to 2018 (GARNET).

INTERVENTIONS

ICI monotherapy or combination therapy in the first-line or later-line setting.

MAIN OUTCOMES AND MEASURES

Cox proportional hazards regression and receiver operating characteristic-area under the curve modeled associations between Serial CTRS and OS.

RESULTS

The study included 1830 patients (RCP discovery, 1171 patients; RCP test, 605 patients; GARNET, 54 patients) with a median (IQR) age of 67 (19-95) years; 1000 participants were male (55%), and 830 were female (45%). Serial CTRS was associated with OS in multivariable analysis controlling for age, sex, programmed death-ligand 1 expression, histologic profile, and tumor volume (hazard ratio [HR] for 10%-point higher probability of 12-month OS, 0.74 [95% CI, 0.70-0.79] for RCP test; 0.45 [95% CI, 0.32-0.65] for GARNET). Serial CTRS outperformed RECIST and TVC in OS risk discrimination, with higher HRs distinguishing low-survival and high-survival groups in RCP test (HR, 6.19; 95% CI, 4.12-9.28) and GARNET (HR, 18.00; 95% CI, 5.40-59.97). Predictive value persisted across programmed death-ligand 1 and RECIST subgroups, including stable disease.

CONCLUSIONS AND RELEVANCE

In this prognostic study of patients with advanced NSCLC receiving ICI treatment, the fully automated biomarker Serial CTRS predicted OS more effectively than resource-intensive RECIST and TVC measurements using the same scans.

Introduction

Accurate prediction of long-term prognosis is a critical unmet need during immune checkpoint inhibitor (ICI) therapy. In routine clinical practice, oncologists often rely on qualitative imaging evaluations and clinical judgment to assess treatment response. Treatment decisions would benefit from objective measures of whether the patient is benefiting from therapy. In clinical trials, efficacy measures such as objective response rate and progression-free survival, assessed using Response Evaluation Criteria in Solid Tumors (RECIST), are widely employed surrogate end points for overall survival (OS), guiding regulatory approvals of oncology therapeutics.1,2,3 Despite their utility, RECIST-based end points exhibit limitations as surrogates of OS.3,4,5,6,7,8,9 RECIST relies on the sum of longest diameters, which fails to capture tumor heterogeneity, irregular shape, mixed responses, immune infiltration, and vascular complexity, compounded by substantial intraobserver and interobserver variability.10,11,12,13,14,15,16 Although enhanced predictive performance has been demonstrated by modeling continuous tumor length or volume changes,17,18,19 these methods are often challenged by the complex response patterns induced by immunotherapies and targeted therapies.20,21,22,23,24,25

Radiomic biomarkers from computed tomography (CT) scans have demonstrated utility in predicting responses to ICIs.26,27,28,29,30,31,32,33 These scans are collected during routine care and for clinical trial RECIST assessments. Traditional radiomic methods extract predefined image features (eg, shape, intensity, and texture) that are then analyzed using machine-learning algorithms.34 In contrast, deep-learning radiomic methods learn relevant imaging features directly from raw images, potentially improving prediction accuracy, generalizability, and efficiency compared with predefined features.26,27,34,35,36 Changes in radiomic features over time have been shown to correlate with both early and long-term therapeutic responses to anticancer therapies across tumor types.28,37,38,39,40,41,42,43,44,45 Radiomic biomarkers that extract features from the full CT image46,47 have the potential to learn additional prognostic features as compared with biomarkers focused only on tumor regions. The objective of the present study was to develop and externally validate a fully automated, deep-learning radiomics-based biomarker using pretreatment and early-treatment routine chest CT images to predict long-term OS in patients with advanced non–small cell lung cancer (NSCLC) treated with an ICI.

Methods

This prognostic study followed the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) + Artificial Intelligence reporting guideline.48 Routine clinical practice (RCP) data collection and deidentification was approved by review boards at each institution. Patients treated with an ICI before diagnosis of advanced NSCLC were excluded.

Data Sources

This analysis was informed by retrospectively collected EHR data from routine clinical practice (often referred to as real-world data) from 2279 adult patients (aged >18 years) with advanced NSCLC treated with an ICI as monotherapy or combination therapy in any line of care at 10 US and European institutions, spanning large academic networks to small community clinics. External validation was conducted using data from a single-group, phase 1 trial of patients with advanced NSCLC receiving single-agent dostarlimab in the second-line or later setting (cohort E of GARNET [NCT02715284]), with rigorous study approval and consent.49 Details of data collection, consent, and RCP dataset construction are provided in the eMethods in Supplement 1.

CT Scans

A model using serial CT scans, Serial CT response score (Serial CTRS) was developed, and baseline (pretreatment) and 12-week follow-up scans were used to derive the Serial CTRS. Twelve weeks was selected to align with the standard clinical decision-making after approximately 2 to 4 treatment cycles and as a common time point for RECIST comparison. Scan selection is detailed in the eMethods in Supplement 1.

Serial CTRS Model Development

A multistage pipeline using 14 424 unannotated, unique, thoracic 3-dimensional (3D) CT images from 1171 patients, retrospectively collected between 2020 and 2024, with different scanner models and acquisition and/or reconstruction settings, was used to develop Serial CTRS (eFigure 1 in Supplement 1). Scans underwent automated quality-control process to ensure correct imaging modality and anatomical region, followed by automated standardized image preprocessing (further details on image preprocessing and model training are included in the eMethods in Supplement 1).

Preprocessed 3D images from paired baseline and 12-week scans and their corresponding paired lesion autosegmentation masks were inputted to a deep-learning feature extractor designed to identify imaging features associated with advanced NSCLC. The model architecture consisted of a custom 3D, hierarchical, convolutional neural network trained on multiple survival end points. To prevent overfitting, training incorporated dropout and augmentation techniques. Extracted imaging features were then input into a Cox proportional hazards model trained on OS. A 6-fold cross-validation strategy optimized the key model parameters. The model output, Serial CTRS, is a continuous 0 to 1 score that represents 12-month OS probability.

Serial CTRS Model Evaluation

Once the optimal model configuration was identified using the RCP discovery dataset, the model was locked and validated using the RCP test and GARNET datasets. For each patient, the final score was calculated as the median of the 6-fold model outputs. The model development team (C.S., T.S., and P.J.) was blinded to the GARNET dataset survival outcomes. The GARNET investigation team (B.K., J.H., and J.V.d.A.) performed the statistical analysis separately without access to the training data or model.

Comparison Models

Serial CTRS performance was compared with RECIST assessments and tumor volume change (TVC) at 12 weeks. Lesion annotation methods are described in the eMethods in Supplement 1. Both RECIST and TVC were not inputs to the Serial CTRS pipeline and were only used for model performance evaluations.

Statistical Analysis

OS was measured from start of ICI treatment; any immortal time bias (because all survived to have at least 1 follow-up scan) affected all patients approximately equally. An OS event was death from any cause, with censoring time as date of last contact.

Associations between Serial CTRS and OS were evaluated using Cox proportional hazards regression models. Hazard ratios (HRs) were estimated from univariable and multivariable models adjusted for covariates such as age, sex, programmed death ligand 1 (PD-L1) tumor proportion score (TPS), baseline tumor volume, and histologic profile. To compare categorical (RECIST version 1.1) with continuous measures (Serial CTRS and TVC), thresholds were identified that allowed the continuous variables to align with the percentages of patients with complete or partial response (combined because of the small number of complete responses), stable disease , and progressive disease per RECIST at week 12. These thresholds were determined and fixed during model development using the RCP discovery dataset and applied to the RCP test and GARNET datasets. Additional comparisons of categorical predictors of OS were performed in subgroups defined by negative (<1%), low (1%-49%), and high (≥50%) categories of tumor PD-L1 expression, and within 12-week RECIST categories.

Model performance was evaluated using the OS concordance index (C-index),50 and receiver operating characteristic-area under the curve for OS at 6 months, 12 months (OS12), and 24 months. In addition, we computed Uno C-index using inverse probability of weighting, with weights derived from a Kaplan-Meier estimator of the censoring distribution, to reduce C-index dependence on censoring.51 Positive predictive value and sensitivity of the high-survival probability group and the negative predictive value (NPV) and specificity of the low-survival probability group were also determined for OS12 predictions.

Results

Patients

The study included 1830 patients with advanced NSCLC, treated with ICI (median [IQR] age, 67 [19-95] years; 1000 [55%] male; 830 [45%] female) (Table). The patient flows of the datasets (RCP discovery, 1171 patients; RCP test, 605 patients; GARNET, 54 patients) are shown in eFigure 2 in Supplement 1. Follow-up scans were acquired after initiation of ICI at a median (IQR) of 77 (58-90) days for RCP discovery, 77 (56-91) days for RCP test, and 83 (79-85) days for GARNET. The median (IQR) follow-up durations (regardless of censoring) and number of OS events (death) were 12.2 (6.0-21.4) months and 474 events for RCP discovery dataset, 9.4 (4.8-18.1) months and 255 events for RCP test dataset, and 12.2 (6.8-23.5) months and 34 events for the GARNET dataset. Of the 1171 RCP discovery cases, 284 (25.4%) were censored for OS earlier than 12 months, the target of the response score. A range of PD-L1 TPS was observed, and PD-L1 was frequently not assessed. The median (IQR) Serial CTRS was 0.79 (0.67-0.86) in the RCP test dataset and 0.76 (0.64-0.83) in the GARNET dataset. Distribution of model scores and the associations with observed OS12 percentage are presented in eFigure 3 in Supplement 1.

Table. Patient Demographics and Disease Characteristicsa.

Demographics and characteristics Participants
RCP discovery dataset (n = 1171) RCP test dataset (n = 605) P value GARNET trial dataset (n = 54) P value
Age, median (IQR), y 67.0 (19-95) 68.0 (32-92) .26 66.5 (45-83) .09
Sex, no. (%) .62 .45
Male 643 (55) 324 (54) 33 (61)
Female 528 (45) 281 (46) 21 (39)
Smoking history, No. (%) .01 <.001
Current or previous 589 (50) 260 (43) 0
Never 66 (6) 34 (6) 0
Unknown 516 (44) 311 (51) 54 (100)
Histologic profile, No. (%)
Adenocarcinoma 811 (69) 449 (74) .01 39 (72) <.001
Squamous cell carcinoma 193 (16) 99 (16) 15 (28)
Unknown 167 (14) 57 (9) 0
EGFR variant status, No. (%)
Positive 86 (7) 35 (6) <.001 1 (2) <.001
Negative 254 (22) 324 (54) 42 (78)
Unknown 831 (71) 246 (41) 11 (20)
ALK variant status, No. (%)
Positive 29 (2) 8 (1) <.001 0 <.001
Negative 217 (19) 326 (54) 41 (76)
Unknown 925 (79) 271 (45) 13 (24)
PD-L1 TPS, No. (%)
Negative (<1%) 283 (24) 152 (25) <.001 21 (39) .003
Low (1%-49%) 262 (22) 155 (26) 14 (26)
High (≥50%) 372 (32) 227 (38) 5 (9)
Unknown 254 (22) 71 (12) 14 (26)
Adrenal metastases, No. (%) 151 (15) 77 (16) .66 6 (11) .60
Bone metastases, No. (%) 222 (22) 100 (20) .64 10 (19) .71
Liver metastases, No. (%) 130 (13) 78 (16) .10 9 (17) .51
ICI line of therapy, No. (%)
First line 678 (58) 389 (64) .004 0 <.001
Second line or greater 493 (42) 216 (36) 54 (100)
ICI therapy regimen, No. (%)
Monotherapy		 728 (62) 298 (49) <.001 54 (100) <.001
ICI plus chemotherapy 348 (30) 268 (44) 0
ICI plus other 95 (8) 39 (6) 0
ICI, No. (%)
Pembrolizumab 730 (62) 429 (71) .004 0 <.001
Nivolumab 285 (24) 109 (18) 0
Durvalumab 80 (7) 35 (6) 0
Atezolizumab 76 (6) 32 (5) 0
Dostarlimab 0 0 54 (100)
PFS, median (IQR), mo 10.1 (3.5-22.0) 7.1 (3.2-16.5) .002 7.0 (2.8-15.1) NA
OS, median (IQR), mo 24.6 (10.5-NR) 20.0 (8.0-64.0) .02 18.9 (9.9-37.8) NA
Duration of follow-up, median (range), mob 12.2 (1-73) 9.4 (1-66) .001 12.2 (2.4-50.0) NA
Scanner type, No. (%)c
CT 12 148 (84) 672 (56) <.001 103 (95) .006
CT/PET 2221 (15) 538 (44) 5 (5)
Unknown 67 (0) 0 0
Slice thickness, No. (%)c
<1.5 mm 4609 (32) 258 (21) <.001 27 (25) <.001
≥1.5 to <3 mm 3079 (21) 209 (17) 19 (18)
≥3 to <5 mm 3985 (28) 443 (37) 20 (19)
≥5 mm 2729 (19) 300 (25) 42 (39)
Unknown 34 (0) 0 0
Pixel spacing, No. (%)c
<0.75 mm 7435 (52) 420 (35) <.001 36 (33) <.001
≥0.75 to <1 mm 6577 (46) 662 (55) 70 (65)
≥1 to <1.25 mm 172 (1) 75 (6) 2 (2)
≥1.25 mm 252 (2) 53 (4) 0
Unknown 0 0 0
Open in a new tab

Abbreviations: ALK, anaplastic lymphoma kinase; CT, computed tomography; EGFR, epidermal growth factor receptor; ICI, immune checkpoint inhibitor; NA, not applicable; NR, not reached; OS, overall survival; PD-L1, programmed death ligand 1; PET, positron emission tomography; PFS, progression-free survival; RCP, real-world; TPS, tumor proportion score.

a

P values were derived from the χ2 test of independence (categorical variables) or t test (continuous variables) and were on the basis of comparison with the RCP discovery dataset. Quartile (25%, median, 75%) PFS and OS are Kaplan-Meier estimates, and duration of follow-up is for both censored follow-up and decedents.

b

Refers to follow-time regardless of censoring. Therefore, a low median does not necessarily indicate a high rate of early censoring.

c

Two scans were performed per patient. Whether or not the 2 scans are performed on a machine with the same characteristics is not summarized, because voxel size is harmonized as part of preprocessing.

Serial CTRS as an Early-Response Biomarker for OS

In the univariable model of the RCP test dataset, each 10%-point higher Serial CTRS was associated with a 27% lower hazard of death (HR, 0.73; 95% CI, 0.69-0.77) and was nearly unchanged (HR, 0.74; 95% CI, 0.70-0.79) after controlling for age, sex, PD-L1 TPS, histologic profile, baseline tumor volume, line of therapy for ICI, and chemotherapy combination (Figure 1A). The C-index was 0.71 (95% CI, 0.67-0.74) for the univariable Serial CTRS model and 0.71 (95% CI, 0.67-0.75) for the multivariable Serial CTRS model. The 24-month Uno C-index of the multivariable model was 0.70 (95% CI, 0.66-0.73).

Figure 1. Overall Survival Associations for Serial Computed Tomography Response Score (CTRS).

Figure 1.

Open in a new tab

Forest plots show overall survival in the real-world (RCP) test dataset (A; n = 605) and GARNET dataset (B; n = 54). Missing covariates, as listed in the Table, were addressed by multiple imputation (programmed death-ligand 1 [PD-L1] tumor proportion score [TPS] was imputed as continuous), as described in the eMethods in Supplement 1. AC indicates adenocarcinoma; HR, hazard ratio; ICI, immune checkpoint inhibitor; and SCC, squamous cell carcinoma.

In the GARNET dataset, Serial CTRS was the only covariate with independent contribution to OS prediction in both the univariable analysis (HR, 0.52; 95% CI, 0.40-0.68) and a multivariable analysis controlling for age, sex, PD-L1 TPS, histologic profile, and baseline tumor volume (HR, 0.45; 95% CI, 0.32-0.65; Figure 1B). The C-index was 0.73 (95% CI, 0.63-0.83) for the univariable Serial CTRS model and 0.75 (95% CI, 0.67-0.83) for the multivariable model. Examining robustness of multivariable model findings in a dataset with a small number of events, the 24-month Uno C-index was 0.76 (95% CI, 0.67-0.85), and leave-one-out sensitivity analysis resulted in Serial CTRS HR no more than 0.02 closer to 1.

Receiver operating characteristic-area under the curve for continuous Serial CTRS predicting OS are reported in the eTable in Supplement 1 for OS at 6 months, OS at 12 months, and OS at 24 months. Estimated area under the curve ranged from 0.74 to 0.76 in the RCP test dataset and from 0.76 to 0.91 in the GARNET dataset.

Comparison of Serial CTRS With RECIST and TVC

To enable a direct comparison with the ability of RECIST to predict OS, Serial CTRS and TVC continuous values were stratified into 3 categories using predetermined thresholds (eFigure 3 in Supplement 1). In both the RCP test and GARNET datasets, Serial CTRS predicted OS better than RECIST and TVC. In the RCP test dataset, the HR for the low vs high survival probability groups was greatest for Serial CTRS (6.19; 95% CI, 4.12-9.28) followed by TVC (4.55; 95% CI, 2.97-6.98) and RECIST (3.79; 95% CI, 2.49-5.77) (Figure 2A-2C). Similarly, in the GARNET dataset, the HR for the low vs high survival probability groups was 18.00 (95% CI, 5.40-59.97) for Serial CTRS vs 12.21 (95% CI, 4.13-36.04) for RECIST and 2.44 (95% CI, 0.95-6.26) for TVC (Figure 2D-2F).

Figure 2. Comparison of Overall Survival (OS) Prediction by Serial Computed Tomography Response Score (CTRS), Response Evaluation Criteria in Solid Tumors (RECIST) Version 1.1, and Tumor Volume Change.

Figure 2.

Open in a new tab

Kaplan-Meier plots show OS stratified by 12-week RECIST response (A and D), percentage change in tumor volume at week 12 (B and E), and Serial CTRS (C and F) for the RCP test (A, B, and C) and GARNET (D, E, and F) datasets. The x-axis is truncated at 24 months, but additional events beyond month 24 were included in the Cox proportional hazards models used to estimate the hazard ratios (HRs) for OS. Serial CTRS and tumor volume change categorized using thresholds identified in the RCP discovery set for direct comparison to 12-week RECIST response. For Serial CTRS, low probability of OS at 12 months is defined as 0 to 0.64, intermediate probability of OS at 12 months is defined as 0.64 to 0.83, and high probability of OS at 12 months is defined as 0.83 to 1. For tumor volume change, low is defined as 34.58% or greater increase, high is defined as 56.50% or greater decrease, and intermediate is defined as being between low and high. C-index indicates concordance index; CR, complete response; PD, progressive disease; PR, partial response; and SD, stable disease.

Serial CTRS also showed strong differentiation when the low or high survival probability group was compared with the intermediate-survival probability group. In the GARNET dataset, the HR for the intermediate vs high survival probability group was 2.91 (95% CI, 1.16-7.31) for Serial CTRS, an improvement over both RECIST (1.34; 95% CI, 0.57-3.13) and TVC (1.00; 95% CI, 0.43-2.34). For the comparison of the low and intermediate survival probability groups, Serial CTRS maintained robust predictive separation, with an HR of 2.82 (95% CI, 2.03-3.94) in the RCP test and 6.19 (95% CI, 2.39-16.01) in the GARNET datasets. As to overall concordance, the Serial CTRS model evaluated on the RCP test dataset had a higher C-index (0.73; 95% CI, 0.69-0.77) than RECIST (0.63; 95% CI, 0.59-0.67) and TVC (0.68; 95% CI, 0.64-0.72). In the GARNET dataset, the C-indices were 0.71 (95% CI, 0.63-0.79) for Serial CTRS, 0.70 (95% CI, 0.62-0.77) for RECIST, and 0.58 (95% CI, 0.46-0.69) for TVC. Binary classification performance, including receiver operating characteristic-area under the curve, positive predictive value, and NPV is presented in the eResults in Supplement 1.

Analysis Within PD-L1 TPS Subgroups

When analyzing the RCP test dataset with both RECIST and PD-L1 measurements by PD-L1 TPS subgroups (n = 168 high, n = 107 low, n = 111 negative), Serial CTRS outperformed both RECIST and TVC in predicting OS in each PD-L1 subgroup (Figure 3A-C for Serial CTRS; eFigure 5 in Supplement 1 for comparisons). In the PD-L1–high subgroup, the C-index was 0.75 (95% CI, 0.69-0.81) for Serial CTRS, 0.62 (95% CI, 0.54-0.70) for RECIST, and 0.67 (95% CI, 0.59-0.75) for TVC. In the PD-L1–low subgroup, these values were 0.72 (95% CI, 0.63-0.81) for Serial CTRS, 0.61 (95% CI, 0.53-0.69) for RECIST, and 0.66 (95% CI, 0.57-0.75) for TVC. Among patients with PD-L1–negative tumors, Serial CTRS maintained its performance, with a C-index of 0.77 (95% CI, 0.69-0.85); the corresponding values were 0.64 (95% CI, 0.57-0.71) for RECIST and 0.70 (95% CI, 0.63-0.77) for TVC.

Figure 3. Overall Survival (OS) Prediction by Serial Computed Tomography Response Score (CTRS) Across Programmed Death-Ligand 1 (PD-L1) and 12-Week Response Evaluation Criteria in Solid Tumors (RECIST) Version 1.1 Subgroups.

Figure 3.

Open in a new tab

Kaplan-Meier plots show OS stratified by 12-week Serial CTRS in subgroups defined by PD-L1 tumor proportion score (TPS) expression level: high (≥50%; A), low (1%-49%; B), and negative (<1%; C), and by 12-week RECIST response: complete response (CR) or partial response (PR) (D), stable disease (SD) (E), and progressive disease (PD) (F). Analysis was performed on the routine clinical practice test dataset. C-index indicates concordance index; and HR, hazard ratio.

Analysis Within 12-Week RECIST Response Groups Including Patients With Stable Disease

In the RCP test dataset, 181 patients had stable disease per RECIST at week 12, with 64 in the high survival probability group, 89 in the intermediate survival group, and 28 in the low survival probability group on the basis of Serial CTRS. The HR for the comparison of low vs high survival probability Serial CTRS was 5.47 (95% CI, 2.75-10.91), showing strong stratification within the stable disease response category. In terms of OS12, the NPV of the low survival probability group was 0.90 (95% CI, 0.68-0.99), with a specificity of 0.28 (95% CI, 0.18-0.41) and C-index was 0.68 (95% CI, 0.61-0.75). Serial CTRS also stratified OS in the subgroups of patients with progressive disease and complete or partial response at week 12 (Figure 3D-F), outperforming TVC in each subgroup (eFigure 6 in Supplement 1).

Discussion

In this prognostic study, we developed and externally validated a fully automated, deep-learning, serial imaging-based biomarker of OS (Serial CTRS) that uses routine thoracic CT scans. Incorporating imaging signals from the entire chest region, Serial CTRS significantly predicted OS probability after adjusting for established prognostic factors and provided superior risk stratification when compared with early-response measures RECIST and TVC. Serial CTRS also stratified OS risk across different PD-L1 expression subgroups with an improved HR between groups, suggesting potential clinical utility beyond standard assessments. Additionally, we found Serial CTRS to enhance risk stratification in patients with stable disease per RECIST at week 12. In this patient subgroup, treatment and clinical trial decisions on the basis of early-response data can be complicated by heterogeneous clinical outcomes.

Most previous studies investigating the prognostic value of serial, CT-based radiomic models used features extracted from tumor regions and typically required manual tumor identification.28,37,38,39,40,41,42,43,44,45 In contrast, Serial CTRS automatically learns features related to OS from the entire chest, including changes in body composition,52 cardiac and vascular regions, and tumor appearance and location. Trebeschi et al46 also investigated a whole-chest serial radiomics model for advanced NSCLC, demonstrating an area under the curve of 0.74 for OS12. The current study provides additional analysis comparing Serial CTRS with RECIST categories and TVC.

Notable strengths of this study are the use of a large and heterogeneous RCP dataset for Serial CTRS development (n = 1171) and testing (n = 605), and external validation using clinical trial data. In fact, the multivariable model for the GARNET dataset showed a greater reduction in the HR for OS with each 10%-point higher Serial CTRS (0.45; 95% CI, 0.32-0.65) than the RCP test dataset (0.74; 95% CI, 0.70-0.79). This may reflect standardization of patient selection, CT acquisition, and assessment timings during clinical trial execution relative to the clinical practice setting. In contrast, centralized PD-L1 assessments did not significantly predict OS in the GARNET dataset (Figure 1), highlighting the potential added value of imaging biomarkers like Serial CTRS over conventional, tissue-based biomarkers.

Because Serial CTRS requires no image annotations, it can eliminate variability in current practice due to subjective manual target lesion selection and diameter measurements.10 In the clinical setting, in which RECIST is not routinely performed, Serial CTRS may provide benefit by identifying patients with high risk, particularly those presenting as stable or mixed response after 6 to 12 weeks of treatment. Additional studies of Serial CTRS patients with high risk are needed to determine effective treatment strategies, which could include increased monitoring or therapy changes. Serial CTRS also identified patients with favorable OS presenting with progressive disease at 12 weeks (eFigure 6 in Supplement 1). Overall in the RCP test data, patients stratified into the serial CTRS low survival probability group had an NPV of 81% for OS12 (ie, 81% did not survive to 12 months). Conversely, the Serial CTRS high survival probability group achieved a positive predictive value of 78% (ie, 78% survived to 12 months) (eResults in Supplement 1). This predictive performance supports future evaluation of Serial CTRS to guide treatment and monitoring strategies (eResults in Supplement 1).

In drug development, RECIST remains important because it provides a standardized and widely accepted assessment of tumor burden changes. Given the improved OS prediction, incorporating Serial CTRS as an end point alongside RECIST could increase confidence, especially when data are immature. Given the modest success rate of oncology trials progressing from phase 2 to phase 3 (approximately 25%-30%) and phase 3 studies achieving their objectives (approximately 60%-70%),53 improved early-stage predictors of long-term outcomes are needed. Notably, only 125 of 392 (32%) oncology drugs approved by the US Food and Drug Administration between 2006 and 2023 ultimately demonstrated a significant OS benefit.3 Improving the prediction of OS is particularly valuable when objective response rate and/or progression-free survival do not align with OS, even with blinded independent central review to minimize imaging-assessment bias.

CT imaging studies at early follow-up may not capture early signals of disease progression and further advances in predicting long-term OS may be achieved by incorporating complementary modalities. Circulating tumor DNA and plasma-based proteomics tested from blood samples have shown association with OS in patients with advanced NSCLC.54,55,56,57 Deep-learning models trained on hematoxylin-eosin–stained whole slides have shown potential in predicting ICI response.58 Combined with Serial CTRS, these modalities could leverage both targeted biological insights and holistic imaging signals to enhance therapy predictions.

Limitations

This study has several limitations. First, the relatively high censoring rate in the RCP discovery used for model development may affect calibration of Serial CTRS survival predictions. Second, like other deep-learning approaches, Serial CTRS currently offers limited interpretability regarding the imaging features contributing to its predictions; ongoing work aims to improve insight into feature importance and to enhance transparency. Third, Serial CTRS extracts features from the chest and upper abdomen, potentially missing prognostic information in other anatomical regions (notably, central nervous system metastases). Fourth, limited sample size, OS censoring, and missing PD-L1 TPS data limit the precision of subgroup analysis in the validation datasets. In particular, estimated HRs for the GARNET data had wide 95% CIs (Figure 2D-F) and were not robust to removal of influential cases. Although C-index and receiver operating characteristic-area under the curve differences were not strictly consistent across subgroups and outcomes, Serial CTRS showed favorable performance across all groups.

Conclusions

In conclusion, Serial CTRS is an externally validated, fully automated, deep-learning, serial imaging–based biomarker that leverages routine CT scans from baseline and early-response follow-up to predict OS more effectively than RECIST and TVC in ICI-treated patients with advanced NSCLC. The automated design of Serial CTRS facilitates future integration into clinical practice and clinical trial workflows. With further validation across therapeutic modalities, Serial CTRS has the potential to enable more accurate, early treatment readouts in both clinical practice and clinical trial settings.

Supplement 1.

eMethods 1. Data sources

eFigure 1. Serial CTRS development pathway

eFigure 2. Patient Flow

eMethods 2. Serial Computed Tomography Response Score (CTRS) Model Development and Evaluation

eResults. Comparison of Serial CTRS With RECIST and TVC: Binary Classification Performance

eFigure 3. Distribution and Calibration of Serial CTRS

eTable. ROC-AUC for Landmark OS Outcomes by Serial CTRS

eFigure 4. ROC-AUC for Landmark OS Outcomes

eFigure 5. Kaplan-Meier Plots for PD-L1 Subgroups

eFigure 6. Kaplan-Meier Plots for 12-Week RECIST Subgroups

jamanetwopen-e2555759-s001.pdf (1.4MB, pdf)
Supplement 2.

Data Sharing Statement

jamanetwopen-e2555759-s002.pdf (13.3KB, pdf)

References

  • 1.Eisenhauer EA, Therasse P, Bogaerts J, et al. New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1). Eur J Cancer. 2009;45(2):228-247. doi: 10.1016/j.ejca.2008.10.026 [DOI] [PubMed] [Google Scholar]
  • 2.US Food and Drug Administration . Clinical trial endpoints for the approval of cancer drugs and biologics: guidance for industry. December 2018. Accessed December 9, 2025. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/clinical-trial-endpoints-approval-cancer-drugs-and-biologics
  • 3.Elbaz J, Haslam A, Prasad V. An empirical analysis of overall survival in drug approvals by the US FDA (2006-2023). Cancer Med. 2024;13(8):e7190. doi: 10.1002/cam4.7190 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Goulart BHL, Mushti SL, Chatterjee S, et al. Correlations of response rate and progression-free survival with overall survival in immunotherapy trials for metastatic non-small-cell lung cancer: an FDA pooled analysis. Lancet Oncol. 2024;25(4):455-462. doi: 10.1016/S1470-2045(24)00040-8 [DOI] [PubMed] [Google Scholar]
  • 5.Shahnam A, Hitchen N, Nindra U, et al. Objective response rate and progression-free survival as surrogates for overall survival treatment effect: a meta-analysis across diverse tumour groups and contemporary therapies. Eur J Cancer. 2024;198:113503. doi: 10.1016/j.ejca.2023.113503 [DOI] [PubMed] [Google Scholar]
  • 6.Merino M, Kasamon Y, Theoret M, Pazdur R, Kluetz P, Gormley N. Irreconcilable differences: the divorce between response rates, progression-free survival, and overall survival. J Clin Oncol. 2023;41(15):2706-2712. doi: 10.1200/JCO.23.00225 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Wang X, Wang X, Hodgson L, et al. Validation of progression-free survival as a surrogate endpoint for overall survival in malignant mesothelioma: analysis of Cancer And Leukemia Group B And North Central Cancer Treatment Group (Alliance) trials. Oncologist. 2017;22(2):189-198. doi: 10.1634/theoncologist.2016-0121 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Blumenthal GM, Karuri SW, Zhang H, et al. Overall response rate, progression-free survival, and overall survival with targeted and standard therapies in advanced non-small-cell lung cancer: US Food and Drug Administration trial-level and patient-level analyses. J Clin Oncol. 2015;33(9):1008-1014. doi: 10.1200/JCO.2014.59.0489 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Prasad V, Kim C, Burotto M, Vandross A. The strength of association between surrogate end points and survival in oncology: a systematic review of trial-level meta-analyses. JAMA Intern Med. 2015;175(8):1389-1398. doi: 10.1001/jamainternmed.2015.2829 [DOI] [PubMed] [Google Scholar]
  • 10.Kuhl CK, Alparslan Y, Schmoee J, et al. Validity of RECIST version 1.1 for response assessment in metastatic cancer: a prospective, multireader study. Radiology. 2019;290(2):349-356. doi: 10.1148/radiol.2018180648 [DOI] [PubMed] [Google Scholar]
  • 11.Cai WL, Hong GB. Quantitative image analysis for evaluation of tumor response in clinical oncology. Chronic Dis Transl Med. 2018;4(1):18-28. doi: 10.1016/j.cdtm.2018.01.002 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Yoon SH, Kim KW, Goo JM, Kim DW, Hahn S. Observer variability in RECIST-based tumour burden measurements: a meta-analysis. Eur J Cancer. 2016;53:5-15. doi: 10.1016/j.ejca.2015.10.014 [DOI] [PubMed] [Google Scholar]
  • 13.Goldmacher GV, Conklin J. The use of tumour volumetrics to assess response to therapy in anticancer clinical trials. Br J Clin Pharmacol. 2012;73(6):846-854. doi: 10.1111/j.1365-2125.2012.04179.x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Muenzel D, Engels HP, Bruegel M, Kehl V, Rummeny EJ, Metz S. Intra- and inter-observer variability in measurement of target lesions: implication on response evaluation according to RECIST 1.1. Radiol Oncol. 2012;46(1):8-18. doi: 10.2478/v10019-012-0009-z [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Oxnard GR, Zhao B, Sima CS, et al. Variability of lung tumor measurements on repeat computed tomography scans taken within 15 minutes. J Clin Oncol. 2011;29(23):3114-3119. doi: 10.1200/JCO.2010.33.7071 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Hobor S, Al Bakir M, Hiley CT, et al. Mixed responses to targeted therapy driven by chromosomal instability through p53 dysfunction and genome doubling. Nat Commun. 2024;15(1):4871. doi: 10.1038/s41467-024-47606-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Bruno R, Marchand M, Yoshida K, et al. Tumor dynamic model-based decision support for phase Ib/III combination studies: a retrospective assessment based on resampling of the phase III study IMpower150. Clin Cancer Res. 2023;29(6):1047-1055. doi: 10.1158/1078-0432.CCR-22-2323 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Struemper H, Rathi C, Muliaditan M, et al. Development of a joint tumor size-overall survival modeling and simulation framework supporting oncology development decision-making. CPT Pharmacometrics Syst Pharmacol. 2025;14(6):1006-1017. doi: 10.1002/psp4.70002 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Maitland ML, Wilkerson J, Karovic S, et al. Enhanced detection of treatment effects on metastatic colorectal cancer with volumetric CT measurements for tumor burden growth rate evaluation. Clin Cancer Res. 2020;26(24):6464-6474. doi: 10.1158/1078-0432.CCR-20-1493 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Fridlyand J, Kaiser LD, Fyfe G. Analysis of tumor burden versus progression-free survival for Phase II decision making. Contemp Clin Trials. 2011;32(3):446-452. doi: 10.1016/j.cct.2011.01.010 [DOI] [PubMed] [Google Scholar]
  • 21.An MW, Dong X, Meyers J, et al. ; Response Evaluation Criteria in Solid Tumors Steering Committee . Evaluating continuous tumor measurement-based metrics as phase II endpoints for predicting overall survival. J Natl Cancer Inst. 2015;107(11):djv239. doi: 10.1093/jnci/djv239 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Kaiser LD. Tumor burden modeling versus progression-free survival for phase II decision making. Clin Cancer Res. 2013;19(2):314-319. doi: 10.1158/1078-0432.CCR-12-2161 [DOI] [PubMed] [Google Scholar]
  • 23.Mandrekar SJ, An MW, Meyers J, Grothey A, Bogaerts J, Sargent DJ. Evaluation of alternate categorical tumor metrics and cut points for response categorization using the RECIST 1.1 data warehouse. J Clin Oncol. 2014;32(8):841-850. doi: 10.1200/JCO.2013.52.3019 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Wolchok JD, Hoos A, O’Day S, et al. Guidelines for the evaluation of immune therapy activity in solid tumors: immune-related response criteria. Clin Cancer Res. 2009;15(23):7412-7420. doi: 10.1158/1078-0432.CCR-09-1624 [DOI] [PubMed] [Google Scholar]
  • 25.Karrison TG, Maitland ML, Stadler WM, Ratain MJ. Design of phase II cancer trials using a continuous endpoint of change in tumor size: application to a study of sorafenib and erlotinib in non small-cell lung cancer. J Natl Cancer Inst. 2007;99(19):1455-1461. doi: 10.1093/jnci/djm158 [DOI] [PubMed] [Google Scholar]
  • 26.Sako C, Duan C, Maresca K, et al. Real-world and clinical trial validation of a deep learning radiomic biomarker for PD-(L)1 immune checkpoint inhibitor response in advanced non-small cell lung cancer. JCO Clin Cancer Inform. 2024;8(8):e2400133. doi: 10.1200/CCI.24.00133 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Saad MB, Hong L, Aminu M, et al. Predicting benefit from immune checkpoint inhibitors in patients with non-small-cell lung cancer by CT-based ensemble deep learning: a retrospective study. Lancet Digit Health. 2023;5(7):e404-e420. doi: 10.1016/S2589-7500(23)00082-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Alilou M, Khorrami M, Prasanna P, et al. A tumor vasculature-based imaging biomarker for predicting response and survival in patients with lung cancer treated with checkpoint inhibitors. Sci Adv. 2022;8(47):eabq4609. doi: 10.1126/sciadv.abq4609 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29.Singh A, Horng H, Roshkovan L, et al. Development of a robust radiomic biomarker of progression-free survival in advanced non-small cell lung cancer patients treated with first-line immunotherapy. Sci Rep. 2022;12(1):9993. doi: 10.1038/s41598-022-14160-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Dercle L, Fronheiser M, Lu L, et al. Identification of non-small cell lung cancer sensitive to systemic cancer therapies using radiomics. Clin Cancer Res. 2020;26(9):2151-2162. doi: 10.1158/1078-0432.CCR-19-2942 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Vaidya P, Bera K, Patil PD, et al. Novel, non-invasive imaging approach to identify patients with advanced non-small cell lung cancer at risk of hyperprogressive disease with immune checkpoint blockade. J Immunother Cancer. 2020;8(2):e001343. doi: 10.1136/jitc-2020-001343 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Trebeschi S, Drago SG, Birkbak NJ, et al. Predicting response to cancer immunotherapy using noninvasive radiomic biomarkers. Ann Oncol. 2019;30(6):998-1004. doi: 10.1093/annonc/mdz108 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.Sun R, Limkin EJ, Vakalopoulou M, et al. A radiomics approach to assess tumour-infiltrating CD8 cells and response to anti-PD-1 or anti-PD-L1 immunotherapy: an imaging biomarker, retrospective multicohort study. Lancet Oncol. 2018;19(9):1180-1191. doi: 10.1016/S1470-2045(18)30413-3 [DOI] [PubMed] [Google Scholar]
  • 34.Scapicchio C, Gabelloni M, Barucci A, Cioni D, Saba L, Neri E. A deep look into radiomics. Radiol Med. 2021;126(10):1296-1311. doi: 10.1007/s11547-021-01389-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Hosny A, Parmar C, Coroller TP, et al. Deep learning for lung cancer prognostication: a retrospective multi-cohort radiomics study. PLoS Med. 2018;15(11):e1002711. doi: 10.1371/journal.pmed.1002711 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36.Mukherjee P, Zhou M, Lee E, et al. A shallow convolutional neural network predicts prognosis of lung cancer patients in multi-institutional computed tomography image datasets. Nat Mach Intell. 2020;2(5):274-282. doi: 10.1038/s42256-020-0173-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.Dercle L, Zhao B, Gönen M, et al. Early readout on overall survival of patients with melanoma treated with immunotherapy using a novel imaging analysis. JAMA Oncol. 2022;8(3):385-392. doi: 10.1001/jamaoncol.2021.6818 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38.Xu Y, Hosny A, Zeleznik R, et al. Deep learning predicts lung cancer treatment response from serial medical imaging. Clin Cancer Res. 2019;25(11):3266-3275. doi: 10.1158/1078-0432.CCR-18-2495 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39.Lin P, Yang PF, Chen S, et al. A delta-radiomics model for preoperative evaluation of neoadjuvant chemotherapy response in high-grade osteosarcoma. Cancer Imaging. 2020;20(1):7. doi: 10.1186/s40644-019-0283-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40.Wang TW, Chao HS, Chiu HY, et al. Evaluating the potential of delta radiomics for assessing tyrosine kinase inhibitor treatment response in non-small cell lung cancer patients. Cancers (Basel). 2023;15(21):5125. doi: 10.3390/cancers15215125 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 41.Gong J, Bao X, Wang T, et al. A short-term follow-up CT based radiomics approach to predict response to immunotherapy in advanced non-small-cell lung cancer. Oncoimmunology. 2022;11(1):2028962. doi: 10.1080/2162402X.2022.2028962 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 42.Liu Y, Wu M, Zhang Y, et al. Imaging biomarkers to predict and evaluate the effectiveness of immunotherapy in advanced non-small-cell lung cancer. Front Oncol. 2021;11:657615. doi: 10.3389/fonc.2021.657615 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43.Cousin F, Louis T, Dheur S, et al. Radiomics and delta-radiomics signatures to predict response and survival in patients with non-small-cell lung cancer treated with immune checkpoint inhibitors. Cancers (Basel). 2023;15(7):1968. doi: 10.3390/cancers15071968 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44.Schutte K, Brulport F, Harguem-Zayani S, et al. An artificial intelligence model predicts the survival of solid tumour patients from imaging and clinical data. Eur J Cancer. 2022;174:90-98. doi: 10.1016/j.ejca.2022.06.055 [DOI] [PubMed] [Google Scholar]
  • 45.Lu L, Dercle L, Zhao B, Schwartz LH. Deep learning for the prediction of early on-treatment response in metastatic colorectal cancer from serial medical imaging. Nat Commun. 2021;12(1):6654. doi: 10.1038/s41467-021-26990-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46.Trebeschi S, Bodalal Z, Boellaard TN, et al. Prognostic value of deep learning-mediated treatment monitoring in lung cancer patients receiving immunotherapy. Front Oncol. 2021;11:609054. doi: 10.3389/fonc.2021.609054 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 47.Torres FS, Akbar S, Raman S, et al. End-to-end non-small-cell lung cancer prognostication using deep learning applied to pretreatment computed tomography. JCO Clin Cancer Inform. 2021;5(5):1141-1150. doi: 10.1200/CCI.21.00096 [DOI] [PubMed] [Google Scholar]
  • 48.Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. doi: 10.1136/bmj-2023-078378 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 49.Moreno V, Roda D, Pikiel J, et al. Safety and efficacy of dostarlimab in patients with recurrent/advanced non-small cell lung cancer: results from cohort E of the phase I GARNET Trial. Clin Lung Cancer. 2022;23(7):e415-e427. doi: 10.1016/j.cllc.2022.05.013 [DOI] [PubMed] [Google Scholar]
  • 50.Harrell FE Jr, Califf RM, Pryor DB, Lee KL, Rosati RA. Evaluating the yield of medical tests. JAMA. 1982;247(18):2543-2546. doi: 10.1001/jama.1982.03320430047030 [DOI] [PubMed] [Google Scholar]
  • 51.Uno H, Cai T, Pencina MJ, D’Agostino RB, Wei LJ. On the C-statistics for evaluating overall adequacy of risk prediction procedures with censored survival data. Stat Med. 2011;30(10):1105-1117. doi: 10.1002/sim.4154 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 52.Chaunzwa TL, Qian JM, Li Q, et al. Body composition in advanced non-small cell lung cancer treated with immunotherapy. JAMA Oncol. 2024;10(6):773-783. doi: 10.1001/jamaoncol.2024.1120 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 53.Seruga B, Ocana A, Amir E, Tannock IF. Failures in phase III: causes and consequences. Clin Cancer Res. 2015;21(20):4552-4560. doi: 10.1158/1078-0432.CCR-15-0124 [DOI] [PubMed] [Google Scholar]
  • 54.Assaf ZJF, Zou W, Fine AD, et al. A longitudinal circulating tumor DNA-based model associated with survival in metastatic non-small-cell lung cancer. Nat Med. 2023;29(4):859-868. doi: 10.1038/s41591-023-02226-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 55.Zou W, Yaung SJ, Fuhlbrück F, et al. ctDNA predicts overall survival in patients with NSCLC treated with PD-L1 blockade or with chemotherapy. JCO Precis Oncol. 2021;5(5):827-838. doi: 10.1200/PO.21.00057 [DOI] [PubMed] [Google Scholar]
  • 56.Frank MS, Andersen CSA, Ahlborn LB, Pallisgaard N, Bodtger U, Gehl J. Circulating tumor DNA monitoring reveals molecular progression before radiologic progression in a real-life cohort of patients with advanced non-small cell lung cancer. Cancer Res Commun. 2022;2(10):1174-1187. doi: 10.1158/2767-9764.CRC-22-0258 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 57.Christopoulos P, Harel M, McGregor K, et al. Plasma proteome-based test for first-line treatment selection in metastatic non-small cell lung cancer. JCO Precis Oncol. 2024;8(8):e2300555. doi: 10.1200/PO.23.00555 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 58.Rakaee M, Tafavvoghi M, Ricciuti B, et al. Deep learning model for predicting immunotherapy response in advanced non-small cell lung cancer. JAMA Oncol. 2025;11(2):109-118. doi: 10.1001/jamaoncol.2024.5356 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplement 1.

eMethods 1. Data sources

eFigure 1. Serial CTRS development pathway

eFigure 2. Patient Flow

eMethods 2. Serial Computed Tomography Response Score (CTRS) Model Development and Evaluation

eResults. Comparison of Serial CTRS With RECIST and TVC: Binary Classification Performance

eFigure 3. Distribution and Calibration of Serial CTRS

eTable. ROC-AUC for Landmark OS Outcomes by Serial CTRS

eFigure 4. ROC-AUC for Landmark OS Outcomes

eFigure 5. Kaplan-Meier Plots for PD-L1 Subgroups

eFigure 6. Kaplan-Meier Plots for 12-Week RECIST Subgroups

jamanetwopen-e2555759-s001.pdf (1.4MB, pdf)
Supplement 2.

Data Sharing Statement

jamanetwopen-e2555759-s002.pdf (13.3KB, pdf)

Articles from JAMA Network Open are provided here courtesy of American Medical Association

RESOURCES